See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260945858

Optimal Dispatch of Photovoltaic Inverters in Residential Distribution Systems

Article in IEEE Transactions on Sustainable Energy · April 2014

DOI: 10.1109/TSTE.2013.2292828

CITATIONS READS
259 288

3 authors, including:

Emiliano Dall'Anese G.B. Giannakis

National Renewable Energy Laboratory University of Minnesota Twin Cities
54 PUBLICATIONS 2,156 CITATIONS 1,249 PUBLICATIONS 72,429 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Emiliano Dall'Anese on 03 November 2014.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 1

Optimal Dispatch of Photovoltaic Inverters

in Residential Distribution Systems
Emiliano Dall’Anese, Member, IEEE, Sairaj V. Dhople, Member, IEEE, and Georgios B. Giannakis, Fellow, IEEE

Abstract—Low-voltage distribution feeders were designed to voltage regulation [4], [8]–[12]. However, such reactive power
sustain unidirectional power flows to residential neighborhoods. control (RPC) strategies typically yield low power factors
The increased penetration of roof-top photovoltaic (PV) systems (PFs) at the feeder input and high network currents, with
has highlighted pressing needs to address power quality and
reliability concerns, especially when PV generation exceeds the the latter translating into power losses and possible conductor
household demand. A systematic method for determining the overheating [13]. To alleviate these concerns, an alternative is
active- and reactive-power set points for PV inverters in resi- to curtail the active power produced by the PV inverters during
dential systems is proposed in this paper, with the objective of peak irradiance hours [13], [14]. For example, in [13], the
optimizing the operation of the distribution feeder and ensuring amount of active power injected by a PV inverter is lowered
voltage regulation. Binary PV-inverter selection variables and
nonlinear power-flow relations render the optimal inverter dis- whenever its terminal voltage magnitude exceeds a prede-
patch problem nonconvex and NP-hard. Nevertheless, sparsity- fined limit. The premise for these active power curtailment
promoting regularization approaches and semidefinite relaxation (APC) strategies is that the resistance-to-reactance ratios in
techniques are leveraged to obtain a computationally feasible low-voltage distribution networks renders the voltages very
convex reformulation. The merits of the proposed approach are sensitive to variations in the active power injections [13], [15].
demonstrated using real-world PV-generation and load-profile
data for an illustrative low-voltage residential distribution system. Of course, pertinent questions in this setup include what is
the optimal amount of power to be curtailed, and by what PV
systems in the network.
Index Terms—Distribution networks, inverter control, photo-
voltaic systems, optimal power flow, voltage regulation, sparsity. A systematic and unified optimal inverter dispatch (OID)
framework is proposed in this paper, with the goal of facil-
itating high PV penetration in existing distribution networks.
The OID task involves solving an optimal power flow (OPF)
I. I NTRODUCTION problem to determine PV-inverter active- and reactive-power
set points, so that the network operation is optimized according
T HE installed residential photovoltaic (PV) capacity in-
creased by 61% in 2012, driven in large by falling
prices, increased consumer awareness, and governmental in-
to well defined criteria (e.g., minimizing power losses), while
ensuring voltage regulation and adhering to other electrical
centives [1]. The proliferation of residential-scale roof-top PV network constraints. The proposed OID framework provides
systems presents a unique set of challenges related to power increased flexibility over RPC or APC alone, by invoking
quality and reliability in low-voltage distribution systems. a joint optimization of active and reactive powers. Indeed,
In particular, overvoltages experienced during periods when while [4], [8]–[10], [12], [13] (and pertinent references therein)
PV generation exceeds the household demand, and voltage demonstrate the virtues of RPC and APC in effecting voltage
sags during rapidly-varying irradiance conditions have become regulation, these strategies are based on local information
pressing concerns [2], [3]. and, therefore, they may not offer system-level optimality
Efforts to ensure reliable operation of the existing distri- guarantees. On the other hand, by promoting network-level
bution system with increased behind-the-meter PV genera- optimization, the OPF-based formulation proposed in this
tion are therefore focused on the possibility of PV inverters work inherently ensures system-level benefits.
providing ancillary services [4]. This setup requires a depar- Unfortunately, due to binary PV-inverter selection vari-
ture from current interconnection standards [5]. Organizations ables and nonlinear power-balance constraints, the formu-
across Industry are endeavoring to address the issue and lated problem is nonconvex and NP-hard. Nevertheless,
bring consistency to grid-interactive controls [6], [7]. For in- a computationally-affordable convex reformulation is de-
stance, by appropriately derating PV inverters, reactive power rived here by leveraging sparsity-promoting regularization
generation/consumption based on monitoring local electrical approaches [16], [17] and semidefinite relaxation tech-
quantities has been recognized as a viable option to effect niques [18]–[21]. Sparsity emerges because the proposed
framework offers the flexibility of controlling only a (small)
subset of PV inverters in the network. This allows one to
Submitted on July 14, 2013; revised on October 13, 2013 and November
21, 2013; accepted November 22, 2013. discard binary optimization variables, and effect inverter selec-
This work was supported by the Institute of Renewable Energy and the tion by using group-Lasso-type regularization functions [16],
Environment (IREE) grant no. RL-0010-13, University of Minnesota. [17]. To cope with nonconvexity due to bilinear power-balance
The authors are with the Digital Technology Center and the Dept. of ECE,
University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455, USA. constraints, the semidefinite relaxation approach [22] proposed
E-mails: {emiliano, sdhople, georgios}@umn.edu in [18], [19] for the OPF problem in balanced transmission
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 2

systems, and further extended to balanced and unbalanced

distribution systems in [20], [23] and [21], respectively, is
employed here too. This approach has the well documented
merit of finding the globally optimal solution of OPF problems
in many practical setups [20], [21]. The resultant relaxed OID
is a semidefinite program (SDP), and it can be efficiently
solved via primal-dual iterations or general purpose SDP
solvers [24], with a significantly lower computational burden
compared to off-the-shelf solvers for mixed-integer nonlinear
programs (MINPs) [25]. Effectively, the OID task can be
implemented in real time, based on short-term forecasts of
Fig. 1. Example of low-voltage residential network with high PV penetration
ambient conditions and loads. Subsequently, state-of-the-art adopted from [3], [13]. Node 0 corresponds to the secondary of the step-down
advanced metering infrastructure can be used to relay set transformer, while set U = {2, 5, 8, 11, 14, 17} collects nodes corresponding
points to inverters [15]. Towards this end, initiatives are to distribution poles. Homes H1 , . . . , H12 are connected to nodes in the set
H = {1, 3, 4, 6, 7, 9, 10, 12, 13, 15, 16, 18}.
currently underway to identify and standardize communication
architectures for next-generation PV inverters [26].
the Frobenius norm. Finally, IN denotes the N × N identity
To summarize, the main contributions of this work are
matrix; and, 0M ×N , 1M ×N the M × N matrices with all
as follows: i) A joint network-wide optimization of PV-
zeroes and ones, respectively.
inverter real and reactive power production improves upon
conventional strategies that solely focus on either real [13] or
reactive power [4], [8]–[11]. ii) Sparsity-promoting regular- B. Distribution-Network Electrical Model
ization approaches are proposed for the OPF problem to offer Consider a low-voltage radial feeder comprising N + 1
PV-inverter selection capabilities. While existing approaches nodes collected in the set N := {0, 1, . . . , N }, and overhead
either require controlling all the PV-inverters, or, assume that lines represented by the set of edges E := {(m, n)} ⊂
nodes providing ancillary services are preselected [4], [8]– N × N . Node 0 is taken to be the secondary of the
[11], [13], [21], the formulated problem identifies optimal step-down transformer. Let U, H ⊂ N collect nodes that
PV inverters for ancillary services provisioning. iii) A third correspond to utility poles, and houses with roof-top PV
contribution is represented by the RPC and APC strategies systems, respectively. For example, consider the residential
outlined in Section IV-C. In fact, the methods in [4], [8]– network in Fig. 1, which is adopted from [3], [13], and is
[10], [13] are based on local information and assume that utilized in the case studies presented subsequently. In this
nodes providing ancillary services are preselected. In contrast, network, one has that U = {2, 5, 8, 11, 14, 17} and H =
the RPC and APC tasks are cast here as instances of the {1, 3, 4, 6, 7, 9, 10, 12, 13, 15, 16, 18} (the latter corresponding
OPF problem, and they offer inverter selection capabilities. to houses H1 , . . . , H12 ).
OPF-type RPC strategies are proposed in [11], [27]; however, The distribution lines (m, n) ∈ E are modeled as π-
inverters providing reactive support are preselected. iv) The equivalent circuits. The series and shunt admittances of line
OID strategy is tested using real-world PV-generation and (m, n) are given by ymn = (Rmn + jωLmn )−1 and ȳmn =
load-profile data for an illustrative low-voltage residential jωCmn , respectively, where Rmn , Lmn , Cmn , and ω denote
distribution system, and its performance is compared with RPC the line resistance, inductance, shunt capacitance, and angular
and APC. frequency [28]. The distribution-network admittance matrix is
The remainder of this manuscript is organized as follows. In denoted by Y ∈ CN +1×N +1 , and its entries are defined as
j∈Nm ȳmj + ymj , if m = n
Section II, the network and PV inverter models are described,
 P

and the RPC, APC, and OID strategies are introduced. Sec- [Y]m,n := −ymn , if (m, n) ∈ E (1)
tion III outlines the proposed OID problem, while Section IV 
0, otherwise
describes its relaxed convex reformulation. Case studies and
implementation details are presented in Section V, while where Nm := {j ∈ N : (m, j) ∈ E} denotes the set of
Section VI concludes the paper and outlines future directions. nodes connected to the m-th one through a distribution line.
Let Vn , In ∈ C denote the phasors for the line-to-ground
voltage and the current injected at node n ∈ N , respectively.
II. P RELIMINARIES AND S YSTEM M ODELS Collecting the currents injected in all nodes in the vector
A. Notation i := [I0 , I1 , . . . , IN ]T ∈ CN +1 , and the node voltages in the
Upper-case (lower-case) boldface letters will be used vector v := [V0 , V1 , . . . , VN ]T ∈ CN +1 , it follows that Ohm’s
for matrices (column vectors); (·)T for transposition; (·)∗ law can be rewritten in matrix-vector form as
complex-conjugate; and, (·)H complex-conjugate transposi- i = Yv. (2)
tion; <{·} and ={·} denote the real and√imaginary parts of
a complex number, respectively; j := −1 the imaginary
unit. Tr(·) the matrix trace; rank(·) the matrix rank; | · | C. Residential PV Inverter and Load Models
denotes the magnitude
√ of a number or the cardinality of a Given prevailing ambient conditions, let P̄h denote the
set; kvk2 := vT v; kvk1 := i |[v]i |; and k · kF stands for
P
maximum active power that can be generated by the PV
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 3

inverter(s) located at node h ∈ H. Henceforth, P̄h is referred

to as the available active power at node h. The |H| × 1 vector
collecting the available active powers {P̄h }h∈H is denoted by
p̄h . Let Ps,h and Qs,h denote the active power injected and
the reactive power generated/consumed by the PV inverter at
node h, respectively. The |H|×1 vectors collecting {Ps,h }h∈H
and {Qs,h }h∈H are denoted by ps and qs , respectively. For
conventional grid-tied residential-scale inverters that do not (a) (b) (c) (d)
offer energy storage capabilities and operate at unity power
factor, it follows that Ps,h = P̄h and Qs,h = 0. Nevertheless, Fig. 2. Operating regions for the PV inverters are shown by the shaded
regions for different strategies: (a) RPC; (b) APC at unity PF; (c) proposed
since strategies where PV inverters are allowed to curtail their OID strategy; and, (d) proposed OID strategy with a constraint on the
power output will be considered in the ensuing subsection, it is minimum PF.
useful to denote the active power curtailed by inverter h ∈ H
as As shown in Fig. 2(c), set FhOID is significantly larger then
Pc,h := P̄h − Ps,h . (3) FhRPC and FhAPC , thus offering improved flexibility when op-
timizing the performance of the residential distribution system.
Furthermore, collect all the curtailed active powers in the |H|× Clearly, the control strategy (6) subsumes RPC and APC.
1 non-negative vector pc . A variant of the OID strategy is when the PV inverters are
A constant P Q model is adopted for the residential loads, required to operate at a sufficiently high PF to limit the circula-
and P`,h and Q`,h denote the residential-load active and tion of reactive power throughout the network (see e.g., [29]).
reactive powers at node h ∈ H. The corresponding |H| × 1 Thus, if the minimum allowed PF is prescribed to be cos θ,
vectors that collect the load active and reactive powers are then the amount of reactive power injectedqor absorbed is
denoted by p` and q` , respectively. constrained by the two bounds: |Qs,h | ≤ Sh2 − Ps,h
2 and
q
cos θ ≤ Ps,h / Ps,h 2 + Q2 .
s,h
D. PV-Inverter Control Strategies The resultant set of feasible inverter operational points is
The maximum reactive power that can be gener- illustrated in Fig. 2(d).
ated/consumed by the h-th inverter is limited by its rated It is worth mentioning that a key implementation task relates
apparent power, which is denoted here as Sh , and the available to formulating active- and reactive-power setpoints for PV
active power P̄h [2], [4]. Specifically, given Sh , the inverter systems composed of multiple inverters, e.g., in microinverter
operating space in the P Q-planep is given by the semicircle
systems. Commercially available microinverter systems usu-
{(P, Q) : 0 ≤ P ≤ Sh , |Q| ≤ Sh2 − P 2 }, whose contour is ally include a power management unit that communicates
represented by a red dotted line in Fig. 2. via power-line communication (see, e.g., [30], [31]). Conse-
If the h-th PV inverter allows reactive power control (RPC), quently, the optimal real- and reactive-power setpoints could
the set of its operating points is given by be relayed from the power management units to the different
q inverters via the power line communication network.
FhRPC := {(Ps,h , Qs,h ) : Ps,h = P̄h , |Qs,h | ≤ Sh2 − P̄h2 } (4)
III. O PTIMAL I NVERTER D ISPATCH P ROBLEM
which indicates that the active power output is the available F ORMULATION
PV power, and the reactive power capability is limited by the
inverter rating [2], [4], [29]. The set FhRPC is represented by Given the operational voltage limits in the residential net-
the vertical segment in Fig. 2(a). When the inverter is not work, the demanded loads p` , q` , and the available PV power
oversized [2], no reactive compensation is possible when P̄h = at the different houses p̄h , the objective is to optimally
Sh ; on the other hand, the entire inverter rating can be utilized dispatch the PV inverters, which amounts to finding the real
to supply reactive power when no active power is produced. and reactive power operating points ps , qs that optimize the
Similarly, the set of feasible operating points when an active operation of the network according to a well defined criterion.
power curtailment (APC) strategy is implemented at inverter To this end, pertinent cost functions are discussed next, while
h [9], [13], [14], is given by the OID problem is formulated in the subsequent Section III-B.

FhAPC := (Ps,h , Qs,h ) : 0 ≤ Ps,h ≤ P̄h , Qs,h = 0


(5) A. Cost Function Formulation
and it is depicted in Fig. 2(b). The cost function to be minimized is defined as
In this paper, an OID strategy whereby PV inverters are
κ(v, i, pc , qs ) := cρ ρ(v, i) + cφ φ(pc ) + cν ν(v) (7)
allowed to adjust both active and reactive powers is proposed.
Consequently, the set of possible operating points is given by where function ρ(v, i) captures real power losses in the
n network; φ(pc ) models possible costs of curtailing active
FhOID := (Ps,h , Qs,h ) : 0 ≤ Ps,h ≤ P̄h , power; and, ν(v) promotes a flat voltage profile. Finally,
cρ > 0, cφ ≥ 0, cν ≥ 0, denote weighting coefficients. The
q o
|Qs,h | ≤ Sh2 − Ps,h
2 . (6)
three components in (7) are substantiated next.
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 4

1) Power losses in the network: Active power losses in the xh = 0, it follows from (11f) and (11g) that Pc,h = 0 and
distribution network are given by [28] Qs,h = 0, thus implying that PV-inverter h is delivering
X the maximum available PV power, P̄h , at unity power factor.
ρ(v, i) := <{Vm Imn∗
} − <{Vn Imn
∗
} (8) Notice that, when xh = 1, the set of admissible setpoints
(m,n)∈E
FhOID for inverter h is specified by constraints (11f)–(11i)
where Imn ∈ C denotes the current flowing on line (m, n). [cf. Figs. 2(c) and 2(d)]. Finally, the constraint on V0 is left
2) Cost associated with active power set points: The cost implicit.
incurred by the utility when inverter h is required to operate The constraint in (11j) offers the possibility of choosing
on a set point different than (P̄h , 0), is modeled without loss (or minimizing) the number of controlled inverters in order
of generality by the following quadratic function to contain the net operational cost of the residential feeder.
X Additionally, alternating between different subsets of K <
φ(pc ) := ah Pc,h
2
+ bh Pc,h (9)
|H| facilitates a more equitable treatment of the inverters (see
h∈H
e.g., [4]). Particularly important is constraint (11e), whose goal
where the choice of the coefficients is based on specific utility- is to ensure the nodal voltages lie between V min and V max
customer prearrangements [29]. For example, if economic (even during intervals with peak solar power generation and
indicators are of interest, the coefficients ah and bh could be low demand). Lastly, (11h)-(11i) follow from the PF constraint
based on the price at which electricity generated by the PV explained in Section II-D.
systems would be sold back to the utility. Similarly, the amount For a fixed assignment of the binary variables {xh }h∈H ,
of curtailed power can be minimized by choosing bh = 1 and (P1) boils down to an instance of the OPF problem. In-
ah = 0, ∀ h ∈ H. deed, similar to various OPF renditions, (P1) is a nonlinear
3) Voltage deviations from average: Voltage deviations nonconvex problem because of the balance equations (11b)-
throughout the network can be minimized by defining (11d), the bilinear terms in (7), and the lower bound in (11e).
Furthermore, the presence of the binary variables {xh }h∈H
v !2
u
uX 1 X renders (P1) an NP-hard MINP, and finding its global opti-
ν(v) := t 2
|Vn | − |Vi | 2 . (10)
N +1 mum requires solving a number of subproblems that increases
n∈N i∈N
exponentially (2|H| ) in the number of PV inverters. In prin-
Function ν(v) encourages flat voltage profiles, since it cap- ciple, off-the-shelf MINP solvers can be employed to find a
tures the distance of the vector
P collecting {|Vn |2 }n∈N from the solution to (P1); see e.g., [25]. However, since these solvers
average vector (1/(N + 1) i∈N |Vi | )1N +1 . It is important
2
are computationally burdensome, they are not suitable for real-
to note that voltage limits will be enforced in the OID problem time feeder optimization [25]. Further, they do not generally
even if deviations from average are not penalized (that is, guarantee global optimality of their solutions, which translates
when cν = 0). Nevertheless, as will be demonstrated in the here to higher power losses and operational costs.
case studies, including this term provides added flexibility in
reducing deviations of the network voltages from the average.
IV. C ONVEX R EFORMULATION
B. OID Problem In this section, a computationally affordable convex refor-
Given the cost function in (7), the OID problem is formu- mulation of (P1) is derived by leveraging sparsity-promoting
lated as follows: regularizations to drop the binary selection variables (Sec-
tion IV-A), and utilizing semidefinite relaxation techniques to
(P1) min κ(v, i, pc , qs ) (11a) address the nonconvexity due to bilinear terms and voltage
{Vn ,In }
{Pc,n ,Qs,n ,xn } lower bounds (Section IV-B).
subject to (2), and
<{Vh Ih∗ } = −P`,h + P̄h − Pc,h ∀ h ∈ H (11b) A. Sparsity-leveraging OID
={Vh Ih∗ } = −Q`,h + Qs,h ∀ h ∈ H (11c) If PV-inverter h is not selected, then it provides the maxi-
Vu Iu∗ =0 ∀ u ∈ U (11d) mum active power P̄h at unity PF; consequently, one clearly
V
min
≤ |Vn | ≤ V max ∀ n ∈ N (11e) has that Pc,h = Qs,h = 0. On the other hand, Pc,h and
Qs,h may be different than zero with OID [cf. Fig. 2].
0 ≤ Pc,h ≤ xh P̄h ∀ h ∈ H (11f)
Supposing that K < |H|, it follows that the 2|H| × 1
Q2s,h ≤ xh (Sh2 − (P̄h − Pc,h ) ) ∀ h ∈ H (11g)
2
real-valued vector [pTc , qTs ]T is group sparse [16]; meaning
Qs,h ≤ xh tan θ(P̄h − Pc,h ) ∀ h ∈ H (11h) that, the 2 × 1 sub-vectors [Pc,h , Qs,h ]T (i.e., the “groups”
−Qs,h ≤ xh tan θ(P̄h − Pc,h ) ∀ h ∈ H (11i) of variables [16]) corresponding to inverters operating at
X point (P̄h , 0) contain all zeroes, whereas [Pc,h , Qs,h ]T 6=
xh ≤ K , {xh } ∈ {0, 1} |H|
(11j) 02×1 for the inverters operated under OID. For example,
h∈H consider the system in Fig. 1, and suppose that only PV-
where xh is a binary variable indicating whether PV-inverter inverters in H10 , H12 are to be controlled. Then, it follows that
h is controlled (xh = 1) or not (xh = 0), and it is assumed [pTc , qTs ]T = [0T18×1 , Pc,10 , Qs,10 , 0T2×1 , Pc,12 , Qs,12 ]T , with
that K < |H| PV-inverters are to be controlled. Clearly, when [Pc,10 , Qs,10 ]T 6= 02×1 and [Pc,12 , Qs,12 ]T 6= 02×1 .
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 5

One major implication of this group sparsity attribute of diag(V) denote the operator that returns a vector collecting
the real-valued vector [pTc , qTs ]T , is that one can discard the the diagonal elements of matrix V. Then, the function ν(v)
binary variables {xh }h∈H , and effect PV inverter selection in (7) can be equivalently expressed in terms of V as ν(V) :=
by leveraging sparsity-promoting regularization techniques. kΠdiag(V)k2 . Next, for each distribution line (m, n) ∈ E,
Among possible candidates, the following group-Lasso-type define the symmetric matrix Lmn := <{ymn }(em −en )(em −
function is well suited for the problem at hand [16], [17]: en )T , and notice that the active power loss on the line
X q (m, n) ∈ E can be re-expressed as a linear function of the
γ(pc , qs ) := λ Pc,h
2 + Q2
s,h (12) matrix V as <{Vm Imn ∗
} −P<{Vn Imn
∗
} = Tr(Lmn V). Thus,
h∈H using matrices Π and L := (m,n)∈E Lmn , the cost function
where λ ≥ 0 is a tuning parameter. Thus, using (12), the OID in (7) becomes
problem in (P1) can be relaxed to:
(P2) min κ(v, i, pc , qs ) + γ(pc , qs ) (13a) κ(V, pc , qs ) = cρ Tr(LV) + cφ φ(pc ) + cν kΠdiag(V)k2 . (15)
{Vn ,In }
{Pc,n ,Qs,n }

subject to (2), (11b) − (11e), and ∀ h ∈ H Notice that for cν ≥ 0, the surrogate cost κ(V, pc , qs ) is
0 ≤ Pc,h ≤ P̄h (13b) convex in the matrix variable V.
Q2s,h ≤ Sh2 − (P̄h − Pc,h )2 (13c) Aiming for an SDP formulation of (P2), Schur’s comple-
Qs,h ≤ tan θ(P̄h − Pc,h ) (13d) ment can be leveraged to convert the non-linear summands
in (12) and (15) to a linear cost over auxiliary optimization
−Qs,h ≤ tan θ(P̄h − Pc,h ). (13e) variables [24] (see Appendix A for details). Specifically, with
Notice the absence of the binary selection variables {xh }h∈H {yh , wh , h = 1, . . . , |H|}h∈H and υ serving as auxiliary
in this formulation. The role of λ is to control the number variables, (P2) can be equivalently reformulated as
of sub-vectors [Pc,h , Qs,h ]T that are set to 02×1 . Particularly,
for λ = 0, all the inverters may curtail active power and X X
(P3) min cρ Tr(LV) + cφ yh + cν υ + λ wh
provide reactive power compensation (that is, K = |H|); and, {Pc,n ,Qs,n }
h∈H h∈H
as λ is increased, the number of inverters participating in {yh ,wh ≥0}
V,υ
OID decreases [16]. Appendix B elaborates further on the PV-
inverter selection capability offered by the regularization func- subject to (13b), (13d) − (13e), (14), and ∀ h ∈ H
υIN +1
 
tion γ(pc , qs ). A generalization of (12) is represented by the Πdiag(V)
0 (16a)
diag(V)T ΠT υ
q
weighted counterpart γw (pc , qs ) := h∈H λh Pc,h + Q2s,h ,
2
P
√
bh Pc,h − yh ah Pc,h
 
where {λh ≥ 0}h∈H substantiate possible preferences to use √ 0 (16b)
(low value of λh ) or not (high value of λh ) specific inverters. ah Pc,h −1
wh 0 Pc,h
 
With this approach, a more equitable treatment of PV inverters
can be facilitated by discouraging the use of inverters whose  0 wh Qs,h   0 (16c)
active and reactive powers are adjusted more frequently. Pc,h Qs,h wh
−Sh2 Qs,h P̄h − Pc,h
 

B. SDP reformulation  Qs,h −1 0 0 (16d)

P̄h − Pc,h 0 −1
Key to deriving an SDP reformulation of (P2) is to ex-
press the active and reactive powers injected per node, and V0 (16e)
voltage magnitudes as linear functions of the outer-product rank(V) = 1 . (16f)
matrix V := vvH . First, let us consider the constraints in
(P2). Let {en }N +1
n=1 denote the canonical basis of R
N +1
,
and define the admittance-related matrix Yn := en en Y per
T The positive semi-definiteness and rank constraints (16e)–(16f)
node n ∈ N . Furthermore, define the Hermitian matrices jointly ensure that for any feasible V ∈ CN +1×N +1 , there
Zn := 21 (Yn + YnH ), Z̄n := 2j (Yn − YnH ), and Mn := en eTn . always exists a vector of voltages v such that V = vvH ,
Then, the nonconvex constraints (11b)-(11d) and (11e) can be and that formulations (P2) and (P3) solve an equivalent
equivalently re-written as (see [19]–[21] for further details): OID problem. For the optima of these two problems, it
holds that Vopt = vopt (vopt )T ; υ opt = kΠdiag(Vopt )k2 ;
Tr(Zh V) = −P`,h + P̄h − Pc,h , ∀h ∈ H (14a) yhopt + zhopt = φ(popt
c ); Tr(LV
opt
) = ρ(vopt , iopt );
P
h∈HP
opt
Tr(Z̄h V) = −Q`,h + Qs,h , ∀h ∈ H (14b) and λ h∈H wh = γ(popt c , qs ), where superscript (·)
opt opt

Tr(Zn V) = 0, Tr(Z̄n V) = 0, ∀n ∈ U (14c) is used to denote the globally optimal solution.

(V ) ≤ Tr(Mn V) ≤ (V
min 2
) ,
max 2
∀n ∈ N (14d) Unfortunately, (P3) is still nonconvex due to the rank-1 con-
straint in (16f). Nevertheless, in the spirit of the semidefinite
which are all linear in the matrix variable V. relaxation technique [22], one can readily obtain the following
Now, let us consider the cost function in (7). Define the (relaxed) SDP reformulation of the OID problem by dropping
projector matrix Π := IN +1 − N1+1 1N +1×1 1TN +1×1 , and let the rank constraint:
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 6

X X
(P4) min cρ Tr(LV) + cφ (yh + zh ) + cν υ + λ wh the following regularization function:
{Pc,n ,Qs,n }
h∈H h∈H
{yh ,wh ≥0} γ 0 (pc , qs ) := γ(pc , qs ) + λp kpc k1 + λq kqs k1 (18)
V,υ
subject to (13b), (13d) − (13e) with λp ≥ 0 and λq ≥ 0 serving as tuning parameters. As in
(14), (16a) − (16d), and V  0. (17) typical (constrained) sparse linear regression problems [33],
the second and third terms in the right hand side of (18)
If the optimal solution V opt
of (P4) has rank 1, then it is also promote entry-wise sparsity in the vectors pc and qs , re-
a globally optimal solution for the nonconvex problems (P3) spectively. Thus, by replacing γ(pc , qs ) with γ 0 (pc , qs ) in
and (P2) [22]. Further, given Vopt , the vectors of voltages√ (P2)-(P4) the following inverter dispatch strategies can be
and injected currents can be computed as vopt := λ1 u1 readily implemented: i) λ > 0, λp = λq = 0: inverters can
and iopt = Yvopt , respectively, where λ1 ∈ R+ is the curtail real power and provide/consume reactive power. ii)
unique non-zero eigenvalue of Vopt , and u1 the corresponding λ = 0, λp > 0, λq = 0, qs = 0: APC-only is implemented
eigenvector. at a subset of PV systems. iii) λ = 0, λp = 0, pc = 0, λq > 0:
A caveat of the semidefinite relaxation technique is that some of the inverters are optimally selected to inject/absorb
matrix V could have rank greater than 1; in this case, reactive power. iv) λ > 0, λp > 0, λq > 0: mixed strategy
the feasible rank-1 approximation of V obtainable via rank whereby all three strategies (RPC, APC, and OID) can be
reduction techniques turns out to be generally suboptimal [22]. implemented.
Fortunately though, when the considered radial residential Compared to the local strategies in [4], [8]–[10], [12],
feeder is single-phase, and the cost (15) is strictly increasing [13], [27], the resultant RPC and APC tasks are cast here as
in the line power flows, Theorem 1 of [20] can be conveniently instances of the OPF problem, thus ensuring system-level opti-
adapted to the problem at hand to show that a rank-1 solution mality guarantees; and, furthermore, through the regularization
of (P4) is always attainable provided that (P2) is feasible, and function (18), they offer the flexibility of selecting the subset
the controlled inverters are sufficiently oversized. Incidentally, of critical PV inverters that must be controlled in order to fulfill
allowing inverters to operate in the region FhOID facilitates optimization objectives and electrical network constraints. An
obtaining a rank-1 solution, since higher reactive powers can OPF-type RPC strategy was previously proposed in [11];
be absorbed (a sufficient condition in [20, Thm. 1] requires however, inverters providing reactive support were preselected.
nodes to be able to absorb a sufficient amount of reactive
power). When the house load is not balanced, the OPF
V. N UMERICAL C ASE S TUDIES
formulation for unbalanced distribution systems in [21] can
be applied to account for discrepancies in nodal voltages and To implement the OID strategy for residential network
conductor currents. optimization, the utility requires: i) the network admittance
Remark 1. The number of controlled PV inverters depends matrix Y; ii) instantaneous available powers {P̄h }; and, iii)
on the parameter λ. Provided (P2) is feasible, there are (at the ratings {Sh }. Once (P4) is solved, the active- and reactive-
least) three viable ways to select λ so that K inverters are power setpoints (i.e., ps , qs ) are relayed to the inverters.
selected: i) First, since both solar irradiation and load typically Consider the distribution network in Fig. 1, which is adopted
manifest daily/seasonal patterns, λ can be selected based on from [3], [13]. The pole-pole distance is set to 50 m, while
historical OID results. ii) In the spirit of the so-called cross-
validation [16], OID problems featuring different values of λ TABLE I
can be solved in parallel. iii) Finally, notice that the optimized S INGLE - PHASE π- MODEL LINE PARAMETERS
values of pc , qs are inherently related to the dual variables R [Ω/km] L [mH/km] C [µF/km]
associated with constraints (14a)–(14b) [16], [17]; then, when Drop line 0.549 0.230 0.055
a primal-dual algorithm is employed to solve (P4), λ can be Pole-pole line 0.270 0.240 0.072
conveniently adjusted during the iterations of this scheme.
Remark 2. A second-order cone program (SOCP) relaxation
1.06
of the OPF problem for radial systems was recently proposed
in [32]. However, sufficient conditions for global optimality 1.05
Voltage magnitude [pu]

of the obtained solution require discarding the upper bounds 1.04

on the real and reactive power injections, which may not be a 1.03

realistic setup for residential distribution systems. In contrast,

1.02
conditions derived in [20] for global optimality of the SDP
No OID
relaxation are typically satisfied in practice. Furthermore, the 1.01
RPC−s1
RPC−s2
virtues of the SDP relaxation in the case of an unbalanced 1 APC−a1
APC−a2
system operation were demonstrated in [21]. 0.99
APC−a3
OID−d1
OID−d2
OID−d3
0.98
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

C. Optimal RPC and APC strategies Node index

It is immediate to formulate optimal RPC and APC strate- Fig. 3. Voltage Profile {|Vn |}n∈N at 12:00 with and without inverter control.
gies in the proposed framework. Towards this end, consider
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 7

the lengths of the drop lines are set to 20 m. The parameters suppose that the active power losses are to be minimized; that
for the admittance matrix are adopted from [3], [13], and is, the weighting coefficients in (15) are cρ = 1, cφ = 0,
summarized in Table I. Voltages V min and V max in (P1) refer and cν = 0. To emphasize the role of the sparsity-promoting
to the minimum and maximum voltage utilization limits for the regularization function, Fig. 4 illustrates the active power
normal operation of residential systems, and they depend on curtailed and the reactive power provided by each inverter for
the specific standard adopted. Since the test system in Fig. 1 λ = 0 (Figs. 4(a) and (b)) and λ = 0.8 (Figs. 4(c) and (d)).
is adopted from [3], [13], these limits are set to 0.917 pu For λ = 0, it is clearly seen that all inverters are controlled; in
and 1.042 pu, respectively in this case study (see page 11 of fact, they all curtail active power from 8:00AM to 6:00PM, and
the CAN3-C235-83 standard). Notice that typical voltages that inject reactive power during the entire interval. Interestingly,
dictate inverters to shut down according to [5] are higher than more active power is curtailed at houses with higher AC ratings
V max . (H3 , H4 , H5 , H7 , H12 ). When λ = 0.8, only 7 inverters are
The optimization package CVX1 along with the interior- controlled, thus corroborating the ability of the regularization
point based solver SeDuMi2 are employed to solve the OID function in (12) in effecting inverter selection. Four remarks
problem in MATLAB. The average computational time required are in order: i) the 7 controlled inverters are the ones located
to solve (P4) was 0.27 s on a machine with a Intel Core i7- far from the transformer; ii) houses at the end of the feeder
2600 CPU @ 3.40GHz. In all the conducted numerical tests, curtail more active power and absorb more reactive power than
the rank of matrix Vopt was always 1, meaning that the glob- the others (thus matching the findings in [3], [13]); iii) all the
ally optimal solution of (P3) was obtained [cf. Section IV-B]. selected inverters absorb reactive power (whereas they inject
All 12 houses feature fixed roof-top PV systems, with a power in Fig. 4(b)); and iv) it is noted that for λ > 0.8, no
DC-AC derating coefficient of 0.77 [34]. The DC ratings of less inverters are selected, meaning that at least 7 inverters
the houses are as follows: 5.52 kW for houses H1 , H9 , H10 must be controlled in order to effect voltage regulation in this
(modeled along the lines of [35]); 5.70 kW (which corresponds particular case study [cf. (14d)]. The resultant voltage profiles
to the average installed rating for residential systems in are illustrated in Fig. 3, where “OID-d1” refers to the case
2011 [1]) for H2 , H6 , H8 , H11 ; and, 9.00 kW for the remaining λ = 0, and “OID-d2” to the case λ = 0.8. Clearly, while
five houses (modeled along the lines of [36]). As suggested the voltage limits are enforced in both cases, a flatter voltage
in [4], it is assumed that the PV inverters are oversized by profile is obtained in the second case.
10% of the resultant AC rating. Further, the minimum PF Next, the OID strategy is compared with: i) RPC without
for the inverters is set to 0.85 [29]. The available active inverter selection (“RPC-r1”); ii) RPC with inverter selection
powers {P̄h } during the day are computed using the System (“RPC-r2”), where λq is such that only 7 inverters provide
Advisor Model (SAM)3 of the National Renewable Energy reactive powers; iii) APC, with K = 12 (“APC-a1”) [3];
Laboratory (NREL), based on the typical meteorological year and, iv) APC with inverter selection (“APC-a2”), where λp
(TMY) data for Minneapolis, MN, during the month of July. in (18) is chosen such that 7 inverters are controlled. It is
The residential load profile is obtained from the Open Energy interesting to note that the minimum PF constraints were not
Info database,4 and the base load experienced in downtown enforced for the RPC strategy; in fact, (P4) is infeasible for
Saint Paul, MN, during the month of July is used for this a minimum PF constraint higher than 0.3. Furthermore, the
test case. To generate different load profiles, the base active case where cφ = 1, bh = cρ = 1 (with all other coefficients
power profile is perturbed using a Gaussian random variable set to 0), and λ, λp chosen such that 7 inverters are selected is
with zero mean and standard deviation 200 W and a PF of also considered. This represents the situation where the utility
0.9 is presumed [3]. Finally, the voltage at the secondary of attempts to minimize the overall power losses in the network,
the transformer is set at 1.02 pu, to ensure a minimum voltage namely the sum of the line losses plus the curtailed power.
magnitude of 0.917 pu at H11 − H12 when the PV inverters The resultant OID and APC strategies are marked as “OID-
do not generate power. d3” and “APC-a3,” respectively. For a fair comparison with
With these data, the peak net active power injection (i.e., RPC, no minimum PF constraints are enforced in OID-d3.
P̄h − P`,h ) occurs approximately at solar noon. Standard As can be seen from Fig. 3, voltage regulation is effected
power-flow computations yield the voltage profile illustrated in all the considered cases, with the flattest voltage profile
in Fig. 3 with a black solid line. Notice that the voltage for APC-a1. However, these strategies yield different active
magnitude towards the end of the feeder (nodes 11-19) exceeds power losses in the network, as well as overall active power
the upper limit; furthermore, the voltage magnitude at houses losses (that is, the power lost in the lines plus the curtailed
H11 , H12 (nodes 17 and 19) is beyond 1.05 pu, which is the power). These two quantities are compared in Fig. 5. When
limit usually set for inverter protection [34]. there is no price associated with the active power curtailed,
Consider then, implementing the proposed OID strategy, and RPC is the one that yields the highest power losses in the
1 [Online]
network (as hinted in [13]); however, APC yields the highest
Available: http://cvxr.com/cvx/
2 [Online] Available: http://sedumi.ie.lehigh.edu/
power losses on the lines when bh = cρ = 1. Interestingly,
3 [Online] Available at https://sam.nrel.gov/. strategy “OID-d3” yields the lowest overall active power
4 “Commercial and Residential Hourly Load Profiles for losses, thus demonstrating the merits of the proposed OID
all TMY3 Locations in the United States,” accessible from approach. On the other hand, “OID-d2” yields a good tradeoff
http://en.openei.org/datasets/node/961. Developed at
the National Renewable Energy Laboratory and made available under the between voltage profile flatness and power loss when bh < cρ .
ODC-BY 1.0 Attribution License. Table II collects the energy loss in the network (in the column
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 8

[W] [VAr]
8:00 1800 8:00
160
9:00 9:00
1600
140
10:00 10:00
1400
120
11:00 11:00

Hour of the day

Hour of the day 12:00
1200
12:00 100

1000
13:00 13:00 80

14:00 800 14:00

60

15:00 600 15:00

40
16:00 16:00
400
20
17:00 17:00
200
0
18:00 18:00
0
H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12

(a) (b)
[W] [VAr]
0
8:00 1800 8:00

9:00 9:00
1600

10:00 10:00 −50

1400
11:00 11:00
Hour of the day

Hour of the day

1200
12:00 12:00
−100
1000
13:00 13:00

14:00 800 14:00

15:00 600 15:00 −150

16:00 16:00
400

17:00 17:00
200 −200
18:00 18:00
0
H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12 H1 H2 H3 H4 H5 H6 H7 H8 H9 H10 H11 H12

(c) (d)

Fig. 4. Dispatched inverters: Curtailed active power (a) and reactive power (b) for λ = 0. Curtailed active power (c) and reactive power (d) for λ = 0.8.

TABLE II
E NERGY LOSS IN THE NETWORK , AND CURTAILED PV ENERGY FOR THE To better highlight the advantages of the proposed OID
SIMULATED DAY method, a long term impact analysis over the course of a
year is also performed. To this end, the hourly profiles for
Network Curtailed Overall
[kWh] [kWh] [kWh] available solar powers are generated using SAM, based on
No control 97.4 0 97.4 the TMY data for Minneapolis, MN, and the hourly load
RPC-r1 15.4 0 15.4 profiles available for the whole year in the Open Energy Info
RPC-r2 18.0 0 18.0 database for the Twin Cities, MN. Table III reports the energy
(cφ = 0)
APC-a1 8.1 118.3 126.4 loss in the network and the energy curtailed over the whole
APC-a2 43.4 75.4 118.8 year. Two setups are considered: i) bh = cφ = 0, cρ 6= 0
OID-d1 4.1 116.6 120.7 (i.e., the active power losses in the network are minimized);
OID-d2 6.5 20.0 26.5
(cφ = cρ ) and, ii) bh = 1, cφ = cρ (i.e., the overall power lost is to
APC-a3 54.4 29.3 83.7 be minimized). In the first setup, it can be clearly seen that
OID-d3 13.7 0.5 14.2 the OID strategy yields the lowest losses in the network. The
proposed scheme outperforms the RPC and APC strategies
also in the second case, since it yields the lowest overall losses.
labeled “Network”), the energy curtailed by the inverters (in To estimate potential economic savings over the course of a
the column labeled “Curtailed”) and the total energy loss (in year, consider using the “average retail price for electricity to
the column labeled “Overall”) for the simulated day. The ultimate customers by end-use sector” available in the monthly
accumulated energy loss in the business-as-usual approach reports of the U.S. Energy Information Administration, for
with no ancillary services is also reported for comparison the year 2012, for the state of Minnesota.5 Let cretail de-
purposes. note the average retail price, and consider solving the OID
To demonstrate the increased flexibility offered by the problem over the whole year in the following two cases:
function cν ν(v) in (7), Fig. 6 depicts voltage profiles at a i) κ(V, pc , qs ) = cretail Tr(LV) (that is, only the economic
few houses during the course of the day for different values losses in the networkP are considered); and, ii) κ(V, pc , qs ) =
of r := cν /cρ , for cφ = 0, cρ = 1, from which it is cretail Tr(LV) + cretail h∈H Pc,h (that is, the economic loss
evident that deviations from the average can be minimized emerges from both the losses in the network and the active
by increasing r. Fig. 6(b) illustrates that the lowest overall power curtailed). Table IV summarizes the overall economic
power losses are obtained for r = 0.5 (demonstrating that r losses over the year. It can be clearly seen that, by enabling a
cannot be increased indiscriminately without considering other
optimization objectives). 5 [Online] Available at: http://www.eia.gov/totalenergy
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 9

10 1.06
r=0
r = 0.5
9 r=2
1.05
8

Voltage magnitude [pu]

Active power loss [kW]

7
1.04

5 1.03

1.02
3

2
1.01

0 1
8 9 10 11 12 13 14 15 16 17 18 8 9 10 11 12 13 14 15 16 17 18
Time of the day Time of the day

(a) (a)

17 17
RPC−r1 No control
16 16
r=0
15
RPC−r2 15 r = 0.5
APC−a1 r=2
14 14
APC−a2

Overall p ower lost [kW]

13 13
APC−a3
Overall power lost [kW]

12 OID−d1 12

11 OID−d2 11

10 OID−d3 10

9 No control 9

8 8

7 7

6 6

5 5

4 4

3 3

2 2

1 1

0 0
8 9 10 11 12 13 14 15 16 17 18 8 9 10 11 12 13 14 15 16 17 18
Time of the day Time of the day

(b) (b)

Fig. 5. (a) Active power loss [kW] on the distribution lines. (b) Sum of the Fig. 6. Voltage magnitude at {Hi , i = 1, 3, 5, 7, 9, 11} (a) and overall power
power lost on the distribution lines and the one curtailed by the inverters. lost (b) for different values of r := cν /cρ . In this setup, cφ = 0, λ = 0.8.

TABLE III TABLE IV

E NERGY LOSS IN THE NETWORK , AND CURTAILED PV ENERGY FOR THE E CONOMIC LOSSES OVER ONE YEAR [$]
WHOLE YEAR
Case RPC APC OID
Network Curtailed Overall cφ = 0, cρ = cretail 540.11 355.20 102.94
[kWh] [kWh] [kWh] bh = 1, cφ = cρ = cretail 540.11 1520.45 499.46
RPC-r1 5040 0 5040
RPC-r2 4710 0 4710
(bh = 0)
APC-a1 965 34170 43820 inverters, in order to enable real-time provisioning of ancillary
APC-a2 3148 23512 26660 services. Future efforts include investigating the impact of
OID-d1 938 34700 35638
OID-d2 913 10862 11775 forecast and communication errors on the inverter dispatch
(cφ = cρ ) task. In addition, the benefits of the proposed OID method
APC-a3 10387 3864 14251 can be effectively communicated to practicing engineers by
OID-d3 4237 143 4380
targeting experimental demonstrations in real-world setups.

joint optimization of the active and reactive power production, A PPENDIX

OID provides the most economic savings. A. Derivation of (P3)
An SDP in standard form involves the minimization of a
VI. C ONCLUDING R EMARKS AND F UTURE D IRECTIONS linear function, subject to linear (in)equalities and linear matrix
A framework to facilitate high PV penetration in existing inequalities (LMIs) [24]. The SDP reformulation of (P2) relies
residential distribution networks is proposed in this paper. on the Schur
  complement. For the symmetric matrix M :=
Through OID, the inverters that have to be controlled in order A B
, where C is invertible, the Schur complement
to enable voltage regulation are identified, and their optimal BT C
active- and reactive-power set points are obtained. The OID of A is given by S = C − BT A−1 B. It follows that M is
problem was cast as an SDP, which is efficiently solvable with positive definite if and only if S and A are positive definite
a computational time on the order of the sub-second. Overall, (or, if and only if S ≥ 0 when A, B, C are scalars); this result
the novel OID framework demonstrates that system-level real- has been leveraged in the derivation of (P3) as detailed next.
time optimization can facilitate the integration of PV systems Aiming for an SDP formulation of (P2), consider intro-
in existing low-voltage distribution systems, and calls for in- ducing the auxiliary variable υ ≥ 0, replacing kΠdiag(V)k2
stituting communication capabilities with residential-scale PV with υ in the cost of (P3), and adding the constraint
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 10

kΠdiag(V)k2 ≤ υ. Then, upon rewriting this constraint as and thresholding vector operation
υ − υ −1 kΠdiag(V)k22 ≥ 0, one can readily obtain (16a) "  −1 #
by using the Schur complement with A = C = υ and opt λ2
xh = I{kch k2 >λ} −η ηQh + I2 ch (21)
B = Πdiag(V). Similarly (16c) can be obtained by setting 2
A = wh I2 , B = [Pc,h , Qs,h ]T , and C = wh ; (16b)
√ with I{A} = 1 if event A is true and zero otherwise, and
can be obtained by setting A = −1, B = ah Pc,h , and
η ∈ R+ the solution of the scalar optimization problem:
C = bh Pc,h −yh ; and (16d) can be obtained by setting A = I2 ,
B = [Qs,h , P̄h − Pc,h ]T , and C = −Sh2 . η

λ2
−1
Notice finally that the positive semi-definiteness and rank min η − cTh ηQh + I2 ch . (22)
η≥0 2 2
constraints (16e)–(16f) jointly ensure that there always exists
a vector of voltages v such that V = vvH for any feasible It can be clearly deduced that xopt
h = 0 when λ ≥ kch k2 ; that
V ∈ CN +1×N +1 (see e.g., [22]). is, inverter h operates at the unitary-PF set point (P̄h , 0). On
the other hand, when [Pc,h , Qs,h ]T 6= 0, the operating point
of inverter h is given by (P̄h − Pc,h , Qs,h ). Equation (21)
B. Soft-thresholding on the inverter set point also explains why, with the decreasing of λ, the number of
To rigorously demonstrate the PV-inverter selection capabil- controlled inverters increases.
ity offered by the proposed relaxed OID problem, results from
duality theory [37] are leveraged next to derive closed-form R EFERENCES
expressions for the optimal inverter set points. [1] L. Sherwood, “U.S. solar market trends 2012,” Jul. 2013, [Online]
Define the real-valued vector xh := P[Pc,h , Q s,h ] , and con-
T Available at http://www.irecusa.org.
sider rewriting (9) as cφ φ(pc , qs ) = h∈H xh Ah xh +bTh xh ,
T [2] E. Liu and J. Bebic, “Distribution system voltage performance analysis
for high-penetration photovoltaics,” Feb. 2008, NREL Technical Moni-
with Ah := diag([cφ ah , 0]) and bh := [cφ bh , 0]T . Notice tor: B. Kroposki. Subcontract Report NREL/SR-581-42298.
further that constraint Q2s,h +(P̄h −Pc,h )2 −Sh2 ≤ 0 [cf. (13c)] [3] R. Tonkoski, R. Turcotte, and T. H. M. El-Fouly, “Impact of high
can be re-expressed in quadratic form as xTh xh +dTh xh +(P̄h2 − PV penetration on voltage profiles in residential neighborhoods,” IEEE
Trans. on Sust. Energy, vol. 3, no. 3, pp. 518–527, Jul. 2012.
Sh2 ) ≤ 0, with dh := [−2P̄h , 0]T . Suppose for simplicity that [4] K. Turitsyn, P. Sulc, S. Backhaus, and M. Chertkov, “Options for control
cν = 0 and no PF constraints are imposed, and consider the of reactive power by distributed photovoltaic generators,” Proc. of the
following relaxed OID problem: IEEE, vol. 99, no. 6, pp. 1063–1073, 2011.
[5] “IEEE 1547 standard for interconnecting distributed resources with
X electric power systems,” [Online]. Available: http://grouper.ieee.org/
(P5) min cρ Tr(LV) + xTh Ah xh + bTh xh + λkx2 k2


groups/scc21/dr shared/.
{xh },V
h∈H [6] California Public Utilities Commission, “Advanced inverter
technologies report,” Jan. 2013, [Online] Available at
subject to (14), V  0, and http://www.cpuc.ca.gov.
0 ≤ Pc,h ≤ P̄h (19a) [7] North American Electric Reliability Corporation, “Special reliability
assessment: Interconnection requirements for variable generation,” Mar.
xTh xh + dH
h xh + (P̄h2 − Sh2 ) ≤ 0. (19b) 2012, [Online] Available at http://www.nerc.com.
[8] P. Carvalho, P. Correia, and L. Ferreira, “Distributed reactive power
Problems (P4) and (P5) are equivalent, and their globally generation control for voltage rise mitigation in distribution networks,”
IEEE Trans. Power Syst., vol. 23, no. 2, pp. 766–772, May 2008.
optimal solution coincide [37]. Further, it can be shown [9] M. A. Mahmud, M. J. Hossain, H. R. Pota, and A. B. M. Nasiruzzaman,
that Slater’s condition holds [19], and thus (P5) has zero “Voltage control of distribution networks with distributed generation
duality gap [37, Ch. 6]. Let {µh }, {µ̄h } denote the multipliers using reactive power compensation,” in Conf. on IEEE Ind. Electr. Soc.,
2011, pp. 985–990.
associated with (14a) and (14b), respectively; {ϕh , ϕ̄h } the [10] A. Cagnano, E. D. Tuglie, M. Liserre, and R. A. Mastromauro, “Online
ones with (14d); and, {γh }, {νh , ν̄h } the ones with the optimal reactive power control strategy of PV inverters,” IEEE Trans.
inverter-related constraints (19a) and (19b), respectively. Fur- on Ind. Electron., vol. 58, no. 10, pp. 4549–4558, Oct. 2011.
[11] M. Farivar, R. Neal, C. Clarke, and S. Low, “Optimal inverter VAR
ther, let L(V, {xh }, {µh , µ̄h , ϕh , ϕ̄h , γh , νh , ν̄h }) denote the control in distribution systems with high PV penetration,” in IEEE PES
Lagrangian of (P5). General Meeting, San Diego, CA, Jul. 2012.
With the optimal dual variables {µopt opt opt opt
h , µ̄h , ϕh , ϕ̄h ,
[12] P. Jahangiri and D. C. Aliprantis, “Distributed Volt/VAr control by PV
opt opt opt inverters,” IEEE Trans. Power Syst., vol. 28, no. 3, pp. 3429–3439, Aug.
γh , νh , ν̄h } known, the Lagrangian optimality condi- 2013.
tion [37, Prop. 6.2.5] asserts that Vopt , {xopt h } can be found as
[13] R. Tonkoski, L. A. C. Lopes, and T. H. M. El-Fouly, “Coordinated
min{xh },V0 L(V, {xh }, {µopt opt opt opt opt opt
h , µ̄h , ϕh , ϕ̄h , γh , νh , ν̄h }).
opt active power curtailment of grid connected PV inverters for overvoltage
prevention,” IEEE Trans. on Sust. Energy, vol. 2, no. 2, pp. 139–147,
Thus, exploiting the decomposability of the Lagrangian, it Apr. 2011.
turns out that the optimal setpoint xh for inverter h is given [14] R. Tonkoski and L. A. C. Lopes, “Impact of active power curtailment on
as the solution of the sub-problem: overvoltage prevention and energy production of PV inverters connected
to low voltage residential feeders,” Renewable Energy, vol. 36, no. 12,
1 T pp. 3566–3574, Dec. 2011.
xopt
h = arg min x Qh x + cTh x + λkxk2 (20) [15] B. A. Robbins, C. N. Hadjicostis, and A. D. Domı́nguez-Garcı́a, “A two-
x 2 stage distributed architecture for voltage control in power distribution
systems,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1470–1482, 2012.
where Qh := 2Ah +2γhopt I2 and ch := bh +µh +ν h +γhopt dh , [16] M. Yuan and Y. Lin, “Model selection and estimation in regression with
with µh := [µopt opt T opt opt
h , −µ̄h ] and ν h := [ν̄h − νh , 0] . Notice
T grouped variables,” J. of the Royal Stat. Soc., vol. 68, pp. 49–67, 2006.
[17] A. T. Puig, A. Wiesel, G. Fleury, and A. O. Hero, “Multidimensional
that Qh  0. Then, from [17, Thm. 1], it follows that the shrinkage-thresholding operator and group LASSO penalties,” IEEE Sig.
optimal set points xh are given by the following shrinkage Proc. Letters, vol. 18, no. 6, pp. 363–366, Jun. 2011.
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY (TO APPEAR) 11

[18] X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming Emiliano Dall’Anese (S’08-M’11) received the
for optimal power flow problems,” Int. J. of Electrical Power & Energy Laurea Triennale (B.Sc Degree) and the Laurea
Systems, vol. 30, no. 6–7, pp. 383–392, Jul.-Sep. 2008. Specialistica (M.Sc Degree) in Telecommunications
[19] J. Lavaei and S. H. Low, “Zero duality gap in optimal power flow Engineering from the University of Padova, Italy,
problem,” IEEE Trans. Power Syst., vol. 1, no. 1, pp. 92–107, Feb. in 2005 and 2007, respectively, and the Ph.D. in
2012. Information Engineering from the Department of In-
[20] A. Y. Lam, B. Zhang, A. Domı́nguez-Garcı́a, and D. Tse, “Optimal formation Engineering, University of Padova, Italy,
distributed voltage regulation in power distribution networks,” 2012, in 2011. From January 2009 to September 2010,
[Online] Available at http://arxiv.org/abs/1204.5226v1. he was a visiting scholar at the Department of
[21] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power Electrical and Computer Engineering, University of
flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, pp. Minnesota, USA. Since January 2011, he has been
1464–1475, Sep. 2013. a postdoctoral associate at the Department of Electrical and Computer
[22] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite Engineering and Digital Technology Center, University of Minnesota, USA.
relaxation of quadratic optimization problems,” IEEE Sig. Proc. Mag., His research interests lie in the areas of power systems, signal processing,
vol. 27, no. 3, pp. 20–34, May 2010. and communications. Current research focuses on energy management in
[23] J. Lavaei, D. Tse, and B. Zhang, “Geometry of power flows and future power systems and grid informatics.
optimization in distribution networks,” in IEEE PES General Meeting.,
San Diego, CA, 2012.
[24] L. Vandenberghe and S. Boyd, “Semidefinite programming,” SIAM
Review, vol. 38, no. 1, pp. 49–95, Mar. 1996.
[25] S. Paudyaly, C. A. Canizares, and K. Bhattacharya, “Three-phase dis-
tribution OPF in smart grids: Optimality versus computational burden,”
in 2nd IEEE PES Intl. Conf. and Exhibition on Innovative Smart Grid
Technologies, Manchester, UK, Dec. 2011.
[26] T. Key and B. Seal, “Inverters to Provide Grid Support (DG and
Storage),” in Proc. of of 5th International Conference on Integration Sairaj V. Dhople (S09-M13) received the B.S.,
of Renewable and Distributed Energy Resources, Dec. 2012. M.S., and Ph.D. degrees in electrical engineering,
[27] S. Bolognani and S. Zampieri, “A distributed control strategy for reactive in 2007, 2009, and 2012, respectively, from the
power compensation in smart microgrids,” IEEE Trans. on Autom. University of Illinois at Urbana-Champaign, IL,
Control, vol. 58, no. 11, pp. 2818–2833, Nov. 2013. USA. He is currently an Assistant Professor in the
[28] J. D. Glover, M. S. Sarma, and T. J. Overbye, Power System Analysis Department of Electrical and Computer Engineer-
and Design. Thomson Learning, 2008. ing at the University of Minnesota (Twin Cities),
[29] M. Braun, J. Künschner, T. Stetz, and B. Engel, “Cost optimal sizing MN, USA, where he is affiliated with the Power
of photovoltaic inverters – influence of of new grid codes and cost and Energy Systems research group. His research
reductions,” in Proc. of 25th Europ. PV Solar Energy Conf. and Exhib., interests include modeling, analysis, and control of
Valencia, Spain, Sep. 2010. power electronics and power systems with a focus
[30] SolarBridge Technologies. (Retrieved, October 2013) Solarbridge power on renewable integration.
manager. [Online]. Available: http://solarbridgetech.com/products/
our-solution/solarbridge-power-manager/
[31] Enphase Energy. (Retrieved, October 2013) Envoy communications
gateway. [Online]. Available: http://enphase.com/products/envoy/
[32] M. Farivar and S. H. Low, “Branch flow model: Relaxations and
convexification (Part I),” IEEE Trans. Power Syst., vol. 28, no. 3, pp.
2554–2564, 2013.
[33] R. Tibshirani, “Regression shrinkage and selection via the Lasso,” J.
Royal Stat. Soc., vol. 58, no. 1, pp. 267–288, 1996.
[34] D. L. King, S. Gonzalez, G. M. Galbraith, and W. E. Boyson, “Perfor-
Georgios B. Giannakis (Fellow’97) received his
mance model for grid-connected photovoltaic inverters,” Sandia National
Diploma in Electrical Engr. from the Ntl. Tech. Univ.
Laboratories, Tech. Rep., September 2007.
of Athens, Greece, 1981. From 1982 to 1986 he
[35] S. T. Cady, D. Mestas, and C. Cirone, “Engineering systems in the
was with the Univ. of Southern California (USC),
rehome: A net-zero, solar-powered house for the U.S. Department of
where he received his MSc. in Electrical Engineer-
Energy’s 2011 Solar Decathlon,” in IEEE Power and Energy Conf. at
ing, 1983, MSc. in Mathematics, 1986, and Ph.D.
Illinois, Univ. Illinois at Urbana-Champaign, 2012.
in Electrical Engr., 1986. Since 1999 he has been a
[36] S. V. Dhople, J. L. Ehlmann, C. J. Murray, S. T. Cady, and P. L.
professor with the Univ. of Minnesota, where he now
Chapman, “Engineering systems in the gable home: A passive, net-
holds an ADC Chair in Wireless Telecommunica-
zero, solar-powered house for the U.S. Department of Energy’s 2009
tions in the ECE Department, and serves as director
Solar Decathlon,” in IEEE Power and Energy Conf. at Illinois, Univ.
of the Digital Technology Center.
Illinois at Urbana-Champaign, 2010.
His general interests span the areas of communications, networking and sta-
[37] D. P. Bertsekas, A. Nedić, and A. Ozdaglar, Convex Analysis and
tistical signal processing - subjects on which he has published more than 360
Optimization. Athena Scientific, 2003.
journal papers, 600 conference papers, 20 book chapters, two edited books and
two research monographs (h-index 105). Current research focuses on sparsity
and big data analytics, wireless cognitive radios, mobile ad hoc networks,
renewable energy, power grid, gene-regulatory, and social networks. He is
the (co-) inventor of 21 patents issued, and the (co-) recipient of 8 best paper
awards from the IEEE Signal Processing (SP) and Communications Societies,
including the G. Marconi Prize Paper Award in Wireless Communications.
He also received Technical Achievement Awards from the SP Society (2000),
from EURASIP (2005), a Young Faculty Teaching Award, and the G. W.
Taylor Award for Distinguished Research from the University of Minnesota.
He is a Fellow of EURASIP, and has served the IEEE in a number of posts,
including that of a Distinguished Lecturer for the IEEE-SP Society.

View publication stats